The Jockey MCP server connects your MCP client to Jockey, a unified agentic system that reasons across your videos and images. Using natural language, you can assemble a collection of videos and images, search across them, ask questions about their contents, and refine results in follow-up turns. Your assistant returns answers as text with clip citations. In clients that render widgets, you can also see a gallery, comparison, timeline, or highlight reel.

Key features:

• Cross-video search: Find clips that match a natural-language query, ranked by relevance.
• Answers with clip citations: Ask a question about your collection and receive a text answer that links to the exact clips it draws from.
• Rendered responses: In clients that render widgets, return results as a clip gallery, comparison, timeline, or highlight reel instead of prose.
• Collection management: Create video collections, add media from a URL, and list their contents without leaving your assistant.
• Multi-turn refinement: Follow up on a previous answer to refine results without restarting.
• Client compatibility: Works with any MCP client that supports remote HTTP servers with OAuth.

Use cases:

• Corpus overview: Summarize themes, subjects, content types, and patterns in a single request.
• Content discovery: Find specific moments in a large collection without manual review.
• Entity tracking: Follow a person, place, brand, or object and build a chronological timeline of appearances.
• Highlight reel assembly: Find and sequence clips around a theme into a curated playlist.
• Library organization: Categorize every video by topic, audience, or any dimension you define.

Key concepts

• Knowledge store: A persistent store of your videos and images plus the understanding the platform derives from them — spatiotemporal context, a typed ontology, and embeddings — that together enable corpus-level reasoning.
• Knowledge store item: An asset added to a knowledge store. The platform processes each item asynchronously. When processing finishes, the item is ready for downstream tasks.
• Multi-turn session: A conversation where each turn builds on previous ones. The conversation retains context across turns, so you can refine prior results without restarting.

Prerequisites

• A TwelveLabs account. If you don’t have one, sign up for a free account.
• An MCP client that supports remote HTTP servers with OAuth. The setup steps below cover Claude and Claude Code. For other clients, follow the client’s documentation.

Set up the MCP server

Follow the steps for your client.

Verify the installation

Ask your assistant to list the tools available from Jockey:

List the tools available from Jockey.

The assistant returns the tools the MCP server exposes for managing knowledge stores, adding media, searching, and answering questions. If the tools don’t appear, see the Troubleshooting section.

Use the MCP server

After the MCP server connects, your assistant can manage knowledge stores, search across videos, and answer questions with clip citations. The examples below show what to ask for each capability.

Build a knowledge store

Ask your assistant to create a knowledge store and add media to it:

Create a knowledge store named “NASA missions” and add these videos: <URL_1>, <URL_2>, <URL_3>.

Each URL should be publicly accessible. For videos, we support files up to 4 GB; for images, the maximum size is 32 MB. The assistant creates the store, adds each URL, and reports the indexing status per item.

Check indexing status

Indexing runs asynchronously and can take several minutes per video. Ask the assistant for the store’s readiness at any time:

Is the “NASA missions” store ready to query?

The assistant reports how many items are still being indexed and how many are ready to be queried.

Generate an overview of a collection

Once your items are ready to query, ask the assistant to summarize the entire collection:

Give me an overview of the “NASA missions” knowledge store. What missions and themes are represented?

Ask for structured JSON when you send the response to another tool:

Return the overview in JSON, including fields for the top themes, notable subjects, and the time range covered.

Search across videos

Find specific moments across every video in a knowledge store:

Find every clip in “NASA missions” that shows an astronaut outside the spacecraft. Show the video and timestamp for each match.

You can narrow the search by modality (visual content, audio, or spoken words) or ask for one clip per video:

Search “NASA missions” by spoken words only. Find every clip where someone says “Earth observation.” Group results by video.

Track entities across videos

Identify and follow specific subjects across an entire library:

List every spacecraft and mission that appears across the videos in “NASA missions”.

Render results as an interactive widget

Ask the assistant to return results as an interactive widget (clip gallery, comparison, timeline, or highlight reel) instead of prose:

Build a 30-second highlight reel of the most visually striking moments in “NASA missions” and show it as a clip timeline.

The model picks each clip, sets its start and end timecodes, decides the order, and adds a short label. The widget shows clip thumbnails, timecodes, source video attribution, and links to each source video.

Refine in a follow-up turn

After the highlight reel above, follow up in the same conversation to refine the clips, the cut, or the layout:

Shorten the reel to 20 seconds and only include moments showing planetary terrain.

Troubleshooting

TwelveLabs tools don’t appear in your assistant

After installing the MCP server, the TwelveLabs tools should appear when you ask your assistant to list them. If they don’t:

• Verify the server is connected. Open your client’s MCP settings and confirm that the jockey server is listed and shows as connected.
• Sign in again. Authorization may have expired. Reconnect and complete the sign-in flow.
• Restart your client. Some clients need a restart after adding a new MCP server.

Authorization fails or the browser shows an error

If the sign-in flow doesn’t complete:

• Confirm your client supports remote MCP with OAuth. Some clients support only local (stdio) servers, or implement OAuth differently than the server expects.
• Restart the sign-in flow from your client, not from the browser.

Searches and questions return no results

If the assistant returns no clips or an empty answer:

• Check that items are ready to query. Ask: “Is the [store name] store ready to query?” Items that are still indexing aren’t searchable.
• Confirm the right knowledge store. Store names aren’t unique. If two stores share a similar name, the assistant may pick the wrong one. Ask the assistant to list your knowledge stores; each has a unique identifier you can reference instead of a name.
• Rephrase the query. Specific terminology may not match what’s in the indexed content. Try a broader or differently-worded prompt.

A widget doesn’t render

The MCP server returns interactive widgets only to clients that support them. Clients without widget support, such as Claude Code, automatically receive a written answer instead. This is expected behavior, not a failure.

Further help

• Check your client’s logs for connection or authentication errors.
• Ask for help in our Discord community.

Single-video tools have moved

The TwelveLabs MCP server for single-video search, analysis, and embeddings is now at mcp-alpic.twelvelabs.io. If your MCP client points at mcp.twelvelabs.io for those tools, update it to the new URL.

For work that spans multiple videos, the Jockey MCP server documented above offers cross-video search, collection summaries, and interactive highlight reels.

Next steps

The guides below cover the TwelveLabs API directly, not the MCP server.

• Quickstart: Build your first knowledge store with the API and learn the underlying workflow.
• Guides: Set up content, configure ingestion, and generate responses.
• Recipes: Common multi-step workflows you can adapt to your own data.